AstroReason-Bench - Stereo Imaging: leaderboard
AstroReason-Bench (Wang et al., Fudan University, 2026) has LLM agents plan satellite operations in a physics-grounded simulator (orbit propagation, slewing, power and onboard storage) through tools and a Python API, all in the same Claude Code agent loop, over a four-day horizon. In the stereo task each target needs a doublet of observations meeting strict azimuth, elevation and timing constraints for 3D reconstruction, with about one satellite per target; the score is the percentage of targets covered by a valid stereo pair, averaged over five generated cases. Higher is better.
Metric: Stereo-pair coverage (%; valid geometric doublets, one run of five procedurally generated cases, Claude Code ReAct harness). Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 12 | #403 |
| 2 | Gemini 3 Flash | 6 | #159 |
| 3 | Claude Sonnet 4.5 | 5 | #244 |
| 4 | DeepSeek V3.1 NEX N1 | 3 | #858 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=astroreason-bench-stereo-imaging · How It Works · Data refreshed daily, snapshot 2026-11-12.